A Comprehensive Analysis of Building Damage in the January 12, 2010

Transcription

1 A Comprehensive Analysis of Building Damage in the January 12, 2010 M w 7 Haiti Earthquake using High-Resolution Satellite and Aerial Imagery Description. A comprehensive analysis of building damage following the 2010 Haiti Earthquake is presented based on the use of high-resolution satellite and aerial imagery. Christina Corbane European Commission/Joint Research Centre TP Ispra Italy Telephone: Fax: christina.corban@jrc.ec.europa.eu Keiko Saito Cambridge Architectural Research Ltd. 25 Gwydir Street, Unit 6 Cambridge, CB1 2LG UK Telephone: , Fax: keiko.saito@carltd.com Luca Dell Oro 1

2 UNOSAT - Operational Satellite Applications Programme United Nations Institute for Training and Research (UNITAR) Palais des Nations CH-1211 Geneva 10 Switzerland Telephone: , Fax: luca.delloro@unitar.org Stuart P.D. Gill The World Bank Group 1818 H Street, NW Washington DC USA Telephone: sgill@worldbank.org Boby Emmanuel Piard Centre National de l Information Géo-Spatiale (CNIGS) 4 Rue Faustin 1er Turgeau, Port-au-Prince Haiti 2

3 Telephone: Charles K. Huyck ImageCat, Inc. 400 Oceangate, Suite 1050 Long Beach, California USA Telephone: , Fax: Thomas Kemper European Commission/Joint Research Centre TP Ispra Italy Telephone: , Fax: Guido Lemoine European Commission/Joint Research Centre TP 267 3

4 21027 Ispra Italy Telephone: , Fax: Robin J.S. Spence Cambridge Architectural Research Ltd. 25 Gwydir Street, Unit 6 Cambridge, CB1 2LG UK Telephone: , Fax: robin.spence@carltd.com Ravi Shankar UNOSAT - Operational Satellite Applications Programme United Nations Institute for Training and Research (UNITAR) Palais des Nations CH-1211 Geneva 10 Switzerland Telephone: , Fax: ravi.santhana@unitar.org 4

5 Olivier Senegas UNOSAT - Operational Satellite Applications Programme United Nations Institute for Training and Research (UNITAR) Palais des Nations CH-1211 Geneva 10 Switzerland Telephone: , Fax: olivier.senegas@unitar.org Francis Ghesquiere The World Bank Group 1818 H Street, NW Washington DC USA Telephone: fghesquiere@worldbank.org David Lallemant The World Bank Group 1818 H Street, NW Washington DC

6 USA Telephone: Galen B. Evans The World Bank Group 1818 H Street, NW Washington DC USA Telephone: gevans@worldbank.org Ross A. Gartley The World Bank Group 1818 H Street, NW Washington DC USA Telephone: rgartley@worldbank.org Joaquin Toro 6

7 The World Bank Group 1818 H Street, NW Washington DC USA Telephone: jtoro@worldbank.org Shubharoop Ghosh ImageCat, Inc. 400 Oceangate, Suite 1050 Long Beach, California USA Telephone: , Fax: sg@imagecatinc.com Walter D. Svekla ImageCat, Inc. 400 Oceangate, Suite 1050 Long Beach, California USA Telephone: , Fax: wsvekla@gmail.com 7

8 Beverley J. Adams ImageCat Ltd. Communications House, 63 Woodfield Lane, Suite G Ashtead, Surrey, KT21 2BT UK Telephone: , Fax: bja@imagecatinc.com Ronald T. Eguchi ImageCat, Inc. 400 Oceangate, Suite 1050 Long Beach, California USA Telephone: , Fax:

9 Abstract The paper provides an account of how three key relief organizations worked together after the devastating Haiti earthquake to produce the first damage assessment based mainly on the use of remotely-sensed imagery. This assessment was jointly conducted by the World Bank (WB), the United Nations Institute for Training and Research (UNITAR) Operational Satellite Applications Programme (UNOSAT), and the European Commission s Joint Research Centre (JRC). This paper discusses the data sources used for the assessment, the methodologies employed to evaluate building damage, and a set of independent studies to validate the final damage results. Finally, a vision of the role of remote sensing technologies in future disasters is presented that serves as a road map for methodological improvements. 9

10 Introduction This paper provides a detailed account of how technology, inspiration and collaboration were used to quickly assess the amount of building damage caused by the 12 January 2010 Haiti earthquake. In less than a minute, this event leveled approximately 20 percent of the buildings in greater Port-au-Prince; killed close to a quarter of a million people; injured as many; and left over a million individuals homeless (Washington Post, 2011). While not considered a great earthquake (by seismological standards), this event will rank as one of the deadliest earthquakes of the 21 st century. This event will also be known as one of the first events where technology (especially high-resolution imagery) was embraced at such a large scale in a real operational sense. Almost from the very onset of the disaster, high-resolution satellite imagery was available to provide the first glimpse of the devastation caused by this earthquake. Days later, very-high resolution aerial imagery was available to provide even more detail on the damage caused in this event. Together, these valuable datasets allowed a small army of remote sensing experts from around the world to provide one of the most comprehensive and rapid assessments of building damage in the last decade. Furthermore, this information was shared with Haitian government officials in relatively short time within two months of the earthquake in the form of a Building Damage Assessment Report in support of the Post-Disaster Needs Assessment (PDNA) and Recovery Framework. This paper discusses how three key international organizations the World Bank (WB), the United Nations Institute for Training and Research (UNITAR) Operational 10

11 Satellite Applications Programme (UNOSAT) and the European Commission s Joint Research Centre (JRC) worked together, in the framework of the joint declaration signed in 2008 between the World Bank, European Commission and the United Nations Development Group on Post-Crisis Assessments and Recovery Planning, to produce this rapid assessment of building damage and how these organizations are continuing this collaboration to improve response protocols for future disasters. It is clear from the Haiti experience that the future of post-disaster damage assessments has changed dramatically and that the bar has been raised for the next event. The paper begins with a brief discussion of the data used in the joint WB- UNOSAT-JRC damage assessment. This discussion is followed by a summary of the different damage evaluation methodologies used by the project teams, focusing on both automated and manual interpretation methods. As part of these discussions, we also present a statistical method of extrapolating building damage totals for severe damage down to moderate or lower damage levels, since the use of nadir aerial imagery often limits what can be seen from the air. We then discuss the various damage maps and products that were jointly produced by the project teams, followed by a detailed discussion of procedures used to validate the final damage totals. We conclude this paper by suggesting some future directions to improve the use of remotely-sensed imagery for comprehensive damage assessment. Datasets: remote sensing and field surveys - One of the legacies from the Haiti earthquake will be the unprecedented amount of free imagery that was made available to the emergency response community. This imagery was critical in identifying impacted 11

12 infrastructure (damaged roads and bridges), destroyed buildings, and the locations of internally displaced people (IDPs). Pre-earthquake imagery was also useful in helping to construct a visual record of the urban landscape before the earthquake destroyed much of the existing building stock. This event was also unusually rich in the types of imagery that were made available to relief organizations and workers. Not only was high-resolution optical imagery available, but even richer datasets (LiDAR) were produced that literally allowed a three-dimensional view or perspective of the earthquake. Because of the significance of the event, these datasets were made available at no cost by many of the data producers and owners. These organizations included the International Charter on Space and Major Disasters, DigitalGlobe, GeoEye, Google, the National Oceanic and Atmospheric Administration (NOAA), the U.S. Geological Survey, Microsoft and the World Bank. Satellite Imagery - The international community benefited from the timely release on 13 January 2010 (i.e., one day after the main shock) of high-resolution satellite data collected by GeoEye with a 50 cm spatial resolution. These data were later complemented by other detailed satellite imagery (e.g., DigitalGlobe) and formed the basis for many of the initial building damage assessments. Especially important in theses releases was the existence of pre-earthquake imagery. High-resolution Aerial Photos - The aerial photos acquired on and after 18 January were of immense value for the comprehensive building damage assessment. The initial satellite imagery had already pointed towards areas of significant damage and total destruction, but these aerial photographs with 15 cm spatial resolution allowed the WB- 12

13 UNOSAT-JRC damage assessment team to further identify damaged buildings. Aerial photos used in the assessment were provided by Google, NOAA-U.S. Geological Survey and the World Bank 1. Compared to the satellite imagery that was so critical in the early damage assessment, the aerial photos increased the level of detail by more than four times, allowing the analysts to extract more information from of the images. Oblique Aerial Imagery - In addition to the above-mentioned data, oblique aerial imagery was made available to the WB-UNOSAT-JRC team by Pictometry, allowing the analysts to access detailed lateral views of the buildings. These data were highly valuable for validation of the joint assessment, as well as to better understand the limitations of images acquired at nadir. Field Survey Data - Extensive field surveys were carried out by members of the WB-UNOSAT-JRC team in close collaboration with CNIGS 2. The data collected through these surveys were used to validate the results of the joint damage assessment. Assessment forms were developed and for each building surveyed, GPS-tagged digital photos were collected and used to assess the damage to that building 3. The validation section of this paper describes this survey in more detail. In addition, 1 The World Bank sponsored the collection of very high resolution optical and LiDAR imagery through the World Bank-ImageCat-Rochester Institute of Technology (WB-IC-RIT) Remote Sensing Mission. A complementary paper in this Special Volume discusses this particular mission in detail. 2 CNIGS is an acronym for Centre National de l Information Géo-Spatiale, a public agency in Haiti. 3 During this deployment, UNOSAT used the ASIGNsign technology by AnsuR Technologies with support from the EC-funded GEO-PICTURES project. 13

14 other surveys (JRC, Betero, Fierro and Perry Engineers, Inc., Stanford University, EERI 4, and EEFIT 5 ) provided valuable ground-based damage datasets that helped to calibrate the results of the joint damage assessment. The use of the ground datasets is discussed in more detail in the section where the methods used to extrapolate damage totals from higher damage grades to lower levels are described. Damage Assessment Methodologies Although each member of the joint assessment team had access to a variety of damage assessment methodologies, it was decided early on that a common methodology based on visual interpretation of damage would be used, which enabled the aggregation of the results. The methodology consisted of using aerial imagery to assign damage grades to individual buildings based on the European Macroseismic Scale (EMS) 1998 (Grünthal, 1998) shown in Figure 1. This scale defines five damage grades: grades 1 to 5 should ideally represent a progressive increase in the strength of shaking for different types of masonry and reinforced-concrete buildings. 4 The Earthquake Engineering Research Institute is a professional organization of engineers and scientists that are dedicated to reducing the effects of earthquake on people and structures. For more information on the organization, please see 5 The Earthquake Engineering Field Investigation Team is an European organization based in the UK with the goal to improve the seismic resistance of structures around the world. 14

15 The detection of damage using remotely-sensed data is done primarily through the analysis of vertical or nadir imagery. Therefore, the types of damage that are recognizable using remotely-sensed data are complete collapses and damage that results in a shifting of building footprints, a lack of definition of the perimeter walls, or obvious building debris. However, some types of failures are not easily detectable using nadir imagery. One example is soft-story failure which was observed in the Haiti earthquake. These failures usually occur when the upper stories of a building collapse directly on to the first or bottom story. These failures are common when the upper stories are constructed to be much stiffer than a more open or softer bottom story. In the nadir images, these types of failures are extremely difficult to identify without the use of oblique imagery or ground survey data. In addition, minor or moderate damage to buildings that do not result in building collapse are also difficult to identify in the nadir images. Accordingly, the damage levels used for marking the individual buildings on the aerial photographs were limited to the higher grades of the EMS-98 scale, e.g. substantial to heavy damage (EMS grade 3), very heavy damage (EMS grade 4), and destruction (EMS grade 5). All buildings that did not exhibit visible damage were labeled with no visible damage (Plate 1). 15

16 Figure 1: Classification of Damage to Masonry and Reinforced Buildings (Taken from EMS, 1998). 16

17 a) No visible damage b) EMS-98 Damage Grade 3 c) EMS-98 Damage Grade 4 d) EMS-98 Damage Grade 5 Plate 1: Example of EMS-98 damage grades as seen in aerial photographs. a) No visible damage. The building with the rusty brown roof appears intact. No debris or evidence of a collapsed structure is observed; b) EMS-98 Grade 3. Substantial damage is observed to building - see debris immediately adjacent to the building, which seems otherwise intact; c) EMS-98 Grade 4. Part of building collapsed, i.e., part of roof or one or more fallen walls, visible 17

18 as bright debris in the adjacent street; and d) EMS-98 Grade 5. Most or all of the building has collapsed - visible as skewered roofs, destroyed walls (loss of shadow) and clearly visible debris surrounding building. Automated Generation of Damage Indicators Although not directly used to establish the final Haiti damage totals, a number of automated methods were employed by team members in order to better understand the relative distribution of damage by region. Post-earthquake damage assessment using imagery has been the subject of numerous methodological studies (Saito et al., 2004, Chiriou, 2005, Huyck et al., 2005, Yamazaki, et al., 2005, Stramondo et al., 2006, Ehrlich et al, 2009). Most of these reviews focus on the use of satellite data, typically from very high resolution optical and SAR sensors. Relatively few studies report on the use of aerial photography, due to the lack of availability of aerial photos taken after earthquakes (Turker et al., 2004). Technical issues highlighted in those studies include a) the need for the appropriate pixel resolution, b) detailed image-to-image registration for pre- and post-event imagery from identical sensor pairs as well as different sensor combinations, i.e., optical-optical (Pesaresi et al, 2007), or optical-sar 6 (Brunner et al, 2010), and c) the limited use of automated processes (Kerle, 2010). Some generic change detection methods (e.g. Bruzzone et al. 2008) show potential, at least in terms of the generation of damage indicators as categorized clusters of change. Such methods can be easily extended to 6 Synthetic Aperture Radar 18

19 work not only on spectral information, but on derived textural (Pesaresi, 2000) or morphological (Dalla Mura et al, 2008) features as well. In operational scenarios, a trade-off must often be made between the sophistication of the image processing workflow and the need for speed in the analysis process, i.e., product delivery. This is particularly true for very high resolution image datasets. In the Haiti earthquake, multiple scenes of pre- and post-event very high resolution (VHR) optical imagery were made available in the first days following the earthquake. The subsequent airborne campaigns delivered approximately 1 TB of digital aerial photography for analysis. An extra complication for the use of automated methods was the absence of a detailed digital elevation model for proper ortho-rectification of the satellite imagery, some of which (e.g., GeoEye imagery over the Goaves area) was acquired at strongly oblique viewing angles. Most of the aerial photography appeared to have not been ortho-rectified, but manually registered to the pre-event satellite image backgrounds. Local displacement of the aerial photography compared to the satellite imagery was clearly evident in the more hilly parts of the scenes. Both the large differences in spatial resolution of the pre-event satellite imagery and the post-event aerial photography and the local displacement precluded the consistent application of automated methods to the full scenes. Locally, though, change detection methods were shown to provide useful information to assist in damage assessment (Tiede et al, 2010). In an effort to generate a damage indicator layer from automatic change detection, applied to satellite imagery pairs to highlight debris patterns of collapsed 19

20 buildings, Gueguen et al. (2010) found the measure to be more effective at identifying improvised shelter locations that were erected on previously open land areas. A number of indicators may be generated from single image sets, for instance, to assist in stratification of the area into classes of interest or to highlight image artifacts representing impact phenomena of interest. In this case, a combination of texture and morphology filters was applied to the full-resolution digital aerial photographs to locate debris and help steer the manual interpretation task. Optimized algorithms were applied overnight, using a 32-CPU cluster machine. However, the final output showed systematic variability over the scene, especially where individual airborne strips were feathered as part of the mosaic generation procedure (Soille, private communication). Texture-based urban density masks, filtered for presence of vegetation, were used to structure the task assignment of the JRC team (Corbane et al., 2010). Visual Interpretation of Building Damage - The decision to join efforts in a collaborative damage assessment exercise was made roughly 2 weeks after the release of the aerial photography and after realization that the analysis of a subset of that data resulted in an order of magnitude more damage locations than in the satellite imagery. By this time, each of the partners had already established their manual interpretation methods, the damage scales to be applied and the analysis environment to support collaborative damage interpretation, and preliminary damage assessments using available satellite imagery were already completed. For instance, whereas both JRC and UNOSAT performed the assessment on a point basis, both used slightly different damage scales initially (EMS-98 and a reverse, reduced version, respectively). The ImageCat/GEO- 20

21 CAN (WB) team had opted for a vector-based or polygon approach (after initially performing a point-basis analysis using satellite imagery) focusing exclusively on the damage classes destroyed (EMS grade 5) and heavily damaged (EMS grade 4). The point-based assessment eventually was adopted by all team members because it allowed for considerably faster damage assignment, and it allowed team members to take advantage of the fact that UNOSAT (working in collaboration with Swisstopo and the University of Zurich) had already started marking points for every building to count the total number of buildings in the study area using pre-event satellite imagery. Although a general agreement was made between the various damage assessment teams, there were some differences in the way the different teams conducted their damage assessments. A common damage scale was agreed upon by all project team members, and the end result of each damage assessment (i.e., point data) was also agreed upon. The following sections describe some of the methodological differences in the various approaches. JRC s damage assessment methodology In the JRC effort, the study region was divided into a series of grid cells, where grid size was based on image resolution and the size of detectable damage artifacts. A stratified sampling approach was used where the strata was defined by the (expected) density of the impact artifacts. For each of the strata, sample clusters were distributed to individual operators. This guaranteed a more or less equal distribution of the initial workload, while the clusters minimized the amount of maneuvering necessary (over the full image set) and maximized the use of any correlation between locally significant damage artifacts. 21

22 Before starting the interpretation process, each analyst went through limited training where examples of previous earthquake damage were presented. Specific attention was given to recognizing the physical changes associated with damaged buildings, the characteristics of the remote sensing sensor (spatial, spectral and temporal resolution), and the environmental conditions at the time of image acquisition (e.g., atmospheric conditions, surface characteristics, seasonal effects, etc.). For the Haiti earthquake, damage patterns differed significantly with type of construction, neighborhood, building density, and terrain slope. The definition of interpretation keys was based on a thorough understanding of the expectations and limitations on the detectability of relevant damage artifacts. Examples of positive and negative detectability were provided to each analyst, i.e., training sets were made available. The analysis of the training results determined whether detectability issues were being correctly understood, or required further illustration. In the course of the interpretation, examples would turn up that illustrated common or special cases of artifacts. Representative samples of such cases were described and added to the training sets so that interpreters became immediately aware of the event-specific cases. All interpretation results were time-stamped, operator-marked and flowed directly into a versioned central repository, so that cross-checks and progress reports could be derived on the fly. Differential progress was used as an indicator to estimate overall damage at an early stage in the interpretation process, or to correct the stratification and sample distribution. 22

23 A separate random sample of grid cells was distributed as a control sample. These samples were interpreted by at least two different interpreters to test for operator bias and further refine training samples. In the Haiti earthquake the most difficult cases, showing up repetitively in quality control, were due to limitations in the vertical perspective inherent to the aerial (or satellite) photography. These included soft-story failures, lateral damage to buildings for which the roofs remained largely intact and which did not show extensive debris around the perimeters, partial collapse of larger building complexes, and very dense informal settlements that were not marked individually, see Plate 2. Most of the false positives were due to buildings that were under construction at the time of the earthquake. Many of these occurred northeast of Port-au-Prince and in Leogane and Jacmel. Plate 2: Examples of false positives and false negatives in the damage assignment by visual interpretation. On the left, buildings under construction in post-earthquake aerial imagery of Croix des Bouquets (north-east of Port-au-Prince) led to commission errors. On the right, an 23

24 oblique Pictometry image shows a soft-story failure of the blue apartment building, which is not visible in the vertical aerial imagery, and hence led to an omission error. Each JRC team worked in a distributed mode, i.e., team members contributed to the damage assessment from different physical locations. No common tools were available across the teams, but each used toolsets that included at least the following characteristics: Synchronized viewing of multiple images of the same area (i.e., side-by-side viewing), in which respective images are properly geo-referenced (i.e., showing the same location). Views of pre- and post-earthquake images in combination with indicator imagery (e.g., sampling strata); A point-and-click function to rapidly assign a classification label to the current image location. Instructions were given to place the mark approximately at the centroid of the building; and Automated and/or manual result synchronization to local files and a central repository. Brunner et al. (2009) proposed addressing such functionalities using Open Source software components. Corbane et al. (2010) also identified the following as being key to a better interpretation exercise: a) automated re-alignment of pre- and post-event imagery that suffers from local displacement, using a correlation approach, with the preevent image as reference; and b) a progress indicator to show sample selection and completion status to the interpreter, and, if relevant, to the rest of the interpreter team(s). 24

25 ImageCat-GEOCAN (WB) damage assessment methodology - A novel aspect of the ImageCat/GEOCAN 7 approach was to use crowd-sourcing as a means of engaging a broader and more global community in the identification of heavily-damaged and/or destroyed buildings in Haiti. Crowd-sourcing or human computation is a relatively new development and certainly a new concept in the area of disaster response. The notion of using a large community of experts to help perform damage assessments in disasters has been introduced by a number of individuals and organizations (Coyle and Meier, 2009). With the Internet becoming the foundation of social networking, it is not unexpected that virtual damage surveys using high-resolution imagery displayed on Google Earth type platforms emerged as a major source of damage information for the Haiti PDNA. After the Haiti earthquake, hundreds of individuals from around the globe sought ways of contributing their expertise, time and knowledge to help support the people of Haiti. ImageCat and the World Bank, working with a network of partners including the Earthquake Engineering Research Institute, reached out to the engineering and scientific community to help in quantifying the extent of damage caused by the Haiti earthquake. In direct response to this event, ImageCat and EERI formed the GEOCAN (Global Earth Observation Catastrophe Assessment Network) community to help the World Bank in its effort to quantify building damage. Using very high resolution aerial imagery, this community was able to identify the number of heavily-damaged and destroyed buildings 7 GEOCAN is an informal organization formed by ImageCat/EERI and the World Bank to rapidly assess damage from global disasters using remotely-sensed data. GEOCAN is an acronym for Global Earth Observation Catastrophe Assessment Network. 25

26 in greater Port-au-Prince. A major product from this effort was the delineation of the preearthquake footprints of these buildings that would later be used by the joint WB- UNOSAT-JRC effort to quantify the amount of building stock that needed to be replaced or repaired. Using crowd-sourcing as the main information technology tool for postdisaster damage assessment, the GEOCAN community was able to deliver its first count of damaged buildings in less than a week. The GEOCAN community consisted of over 600 experts from 23 different countries with over 60 major universities represented, about 20 government agencies and non-profit organizations, and over 50 private companies 8. Current efforts by the World Bank, EERI and ImageCat are ongoing to institutionalize the GEOCAN community as a permanent tool in the World Bank damage assessment toolbox. Extrapolation of Severe Damage States to Lower Damage Levels Because damage at lower levels was difficult to identify using only nadir imagery, a set of statistical models were developed that helped to extrapolate the results of the aerial damage surveys (Grades 4 and 5) to lower damage levels. These models were based on the results of several field surveys conducted after the earthquake in the form of damage distribution matrices so that the relative percentage of buildings in each damage state or grade could be presented for each land-use type. Normally, ground shaking intensity is also added as an independent parameter; however, in the Haiti earthquake, most areas with earthquake damage were rated as having at least a Modified 8 See for a complete list of organizations that participated in the GEOCAN Haiti Initiative. 26

27 Mercalli Intensity (MMI) 9 of 9. That is, all areas experienced significant levels of ground motion (McCann and Mora, in preparation). Damage levels should also be a function of building construction type or design. However, because little or no seismic design considerations were employed in the building of structures in Haiti, all buildings were highly vulnerable to the effects of earthquakes. Therefore, construction type was not a major differentiator in determining whether a building experienced damage or not. There is, however, strong evidence that local soil conditions (including slope gradients) was a major factor in causing damage to buildings. The ImageCat/GEOCAN team worked with various field investigation teams in constructing the damage models used to scale the aerial damage results to lower damage levels. The groups that contributed directly to the ImageCat/GEOCAN effort by providing ground survey data included: Cambridge Architectural Research (CAR) Ltd. Betero, Fierro and Perry (BFP) Engineers, Inc. Stanford University and Pacific Earthquake Engineering Research (PEER) team. In each of these cases, the ImageCat/GEOCAN project team had several sources from which to create these damage distributions. For the commercial/downtown/industrial land-use classes, it was decided that using the field observations from the BFP and Stanford/PEER surveys provided the best source of information from which to judge 9 See for a description of the Modified Mercalli Intensity scale. 27

28 damage for this land-use category. Plate 3 provides a sample of the data used by these teams to determine the damage grade for each building. Initially, a six-grid area was targeted for this ground assessment. Plate 3 highlights Grid 4 with which the image is associated with. Block 7 is a further breakdown of the Grid 4 study area. The footprint boundaries presented in yellow and red are the results of the GEOCAN damage assessment. The larger grid system shown in Block 7 was selected because it represented one of the hardest hit areas in this earthquake, and because it is an area which the survey team had visited in an earlier deployment, thus the team could perform this survey relatively quickly. 28

29 N Each Grid Cell Is 0.5 km by 0.5 km. Six grid cell area surveyed by Eduardo Fierro (BFP, 2010) Plate 3: Plan View of Damage Assignments Completed by BFP for Central Port-au-Prince (PAP) (BFP, 2010) For the other land-use types, the ImageCat/GEOCAN team used the damage distribution results from the CAR Pictometry analysis (Spence and Saito, 2010). Because of access issues (e.g., tall walls), it was felt that using oblique imagery (Pictometry) to determine the damage levels in residential areas was the best source of information for this land-use type. The Pictometry Online system was used extensively by all members of the WB- UNOSAT-JRC team to evaluate building damage within PaP. Access to these data was 29

30 critical in determining damage in targeted areas within PaP. Because it was not easy to get to many areas via field surveys, the Pictometry online system allowed the WB- UNOSAT-JRC project team to remotely survey these harder-to-reach areas. This facilitated a more comprehensive evaluation, although the level of detail for any particular building was not as good as having the results of the field surveys. To estimate building counts for these lower damage categories or grades, the WB- UNOSAT-JRC team used a series of simple equations that calculated the number of buildings in EMS grades 1, 2 and 3 based on the proportion of buildings in each of these categories and the number of observed EMS grade 4 and 5 buildings. In addition, landuse type was a key factor in selecting the proper equation. To illustrate this process, the equation below is provided for grade 3 buildings. No. of Grade 3 buildings = p 3 (p 4 + p 5 ) 1 (n 4 + n 5 ) 1 where p 3 is the percentage (in terms of a ratio) of grade 3 buildings for commercial, downtown, and industrial buildings, p 4 is the percentage for grade 4, p 5 is the percentage for grade 5, n 4 is the number of counted grade 4 buildings from the aerial damage survey, and n 5 is the number of counted grade 5 buildings from the aerial survey. Note that the percentage of buildings in each damage grade was determined from a statistical analysis of the ground-based survey data discussed above and an analysis of the Pictometry data. The procedure was applied separately for each land-use area. The complete set of ratios for all land-use types and communes is contained in World Bank/ImageCat report which is in review (World Bank/GFDRR/ImageCat, in review). These ratios were used in 30

31 estimating the total number of damaged buildings in each of the five damage grades which is reflected in a subsequent section. Damage Validation Studies To analyze the accuracy of the joint damage assessment results described in previous sections, several independent validation datasets were created either from ground surveys or from high-resolution oblique aerial photographs. The following sections describe these validation efforts. Ground Survey Data - One month after the earthquake, JRC, the World Bank and UNOSAT deployed teams to carry out building-by-building ground surveys in the greater Port-au-Prince area including Carrefour, Petionville, and Delmas, as well as in Gressier and Leogane. There were no surveys done in Jacmel or Petite Goave or Grande Goave. The National Geospatial Capacity (Centre National de l Information Géo-Spatiale CNIGS) of Haiti subsequently took over the ground survey work. CNIGS carried out ground surveys for more than 6,000 buildings in the greater Port au Prince area between 8 th March and 7 th May The breakdown of the number of buildings assessed can be found in Table 1. The team also took geo-tagged photographs of the buildings and for some, assessed the damage to the building using these photos. Damage grades were assigned to the buildings using EMS-98 damage scale. Other attributes, such as the number of stories as well as the construction type (reinforced concrete or unreinforced masonry) and building use type (residential, commercial etc) were also recorded. 31

32 Table 1: Breakdown of the Number of Buildings Surveyed by JRC-UNOSAT-WB-CNIGS on the Ground by Commune using Ground Photos Commune Number of buildings assessed using aerial photos/satellite images by WB-UNOSAT-JRC teams Number of buildings surveyed on the ground Croix des Bouques 3, Cite Soleil 6, Carrefour 69,384 1,305 Leogane 23, Delmas 64, Tabarre 3, Petion Ville 10, Gressier 1, Port au Prince 92,432 2,607 TOTAL 276,124 6,492 32

33 Accuracy assessment - Of the 6,492 buildings surveyed on the ground by the JRC-UNOSAT-WB-CNIGS team, only 4,329 of them corresponded to the buildings included in the remote sensing damage assessment. Table 2 shows a comparison of the remote sensing results to the ground survey data. When the comparison is assessed using the original four damage categories in the ground survey (EMS-98 grades 0 to 2, grade 3, grade 4 and grade 5), the overall accuracy of the remote sensing results is 61%. However, when the damage grades are aggregated into only three categories (i.e., grades 3 or less, 4 and 5), as shown in Table 2, the overall accuracy increases to 73%. The commission and omission errors suggest that there is a lot of confusion even for grade 5, i.e., destroyed buildings. Further analysis of the errors has been undertaken and these results show that 20% of the total error, which is directly attributed to the interpreter, can be avoided through better damage assessment protocols (Shankar, 2010). Table 2: Confusion Matrix Comparing the Ground Survey Results and the Aerial Photograph Interpretation of 4329 Buildings in Port-Au-Prince and Surrounding Areas (Table taken and modified from Shankar, 2010) Aerial Photo interpretation by WB-UNOSAT-JRC D D4 D5 Total Producer s accuracy Omission error Ground Survey D D % 20% 30% 70% 33

34 JRC- UNOSAT- D % 53% WB Total User s accuracy Commission 91% 14% 40% error 9% 86% 60% Kappa value: 0.31 Baseline Data Quality Issues - When comparing the ground-based validation data to the damage assessment results derived from the remotely-sensed data, it became clear that there were issues related to the baseline data, i.e., the building points. In some cases, the building points marked by the interpreter were located outside the building boundary. In others, there was more than one building point marked inside a building s boundary. Reducing these errors would be important in order to improve the accuracy of the damage assessment. In one area, a very high accuracy percentage (82%) was achieved using the results of remote sensing assessment; this is attributed to the availability of pre-event building point data. Validation of GEOCAN Assessment using Pictometry Data - Pictometry data consists of oblique aerial photographs taken from four directions as well as nadir. The online user interface allowed access to seamless oblique images for each direction with a spatial resolution of less than 15 cm. Since Pictometry images are oblique, the facades of the buildings are visible, which is an advantage for damage assessment. 34

35 With the lack of ground surveys to validate the damage assessment results derived from the imagery analysis, CAR performed an independent damage assessment using the Pictometry data taken in late January / early February 2010, courtesy of Pictometry, for the purpose of validating the GEOCAN results. Because this validation work had to be completed quickly, a decision was made to produce results for only a sample of sites. Stratified random sampling based on dominant land-use type and neighborhood damage levels was adopted in selecting the locations sampled within Port-au-Prince. Sixty locations in total were selected at equidistance from each other, with approximately an equal number of survey locations chosen for the different land-use types. In each location, about 20 consecutive buildings were evaluated and EMS-98 damage grades were assigned, as well as the building type (construction), number of stories, building use type (residential or commercial) and other comments. All locations were selected on the basis that they could be easily accessed from the main street, in case a ground survey was to be carried out. Subsequently in April 2010, the Earthquake Engineering Field Investigation Team (EEFIT) deployed to Port-au-Prince and undertook ground surveys in 8 of the 60 locations (technical report pending). In total, 1,247 buildings were evaluated using the Pictometry data. The results of the Pictometry damage assessment are presented in Table 3. 35

36 Table 3: Breakdown of Number of Buildings by Damage Grade and Land-Use Class using Pictometry Data Count Percentage % Commercial Downtown Residential Shanty Total Commercial (n=380) Downtown (n=199) Residential (n= 308) Shanty (n=354) D % 16.1% 12.7% 11.9% D % 12.1% 6.5% 9.9% D % 5.0% 10.1% 11.9% D % 12.1% 7.8% 12.1% nvd % 54.8% 63.0% 54.2% total % 100.0% 100.0% 100.0% D4+D5 33.2% 28.1% 19.2% 21.8% * Note: nvd no visible damage The proportion of buildings in Damage grades 4 and 5 is not dissimilar across all four land-use classes. The highest damage rate (33.2% grade 4 or 5) was found in the commercial area; the downtown area (28.2%) was also comparatively badly damaged. Damage proportions in the residential and shanty areas were significantly lower (19.2% and 21.8%, respectively). As with the joint damage assessment using nadir images, a damage grade called NVD (no visible damage) was used in cases where there was no apparent visible damage to the building or where the aerial view was obscured, for instance, by a tree. The Pictometry survey was able to clearly distinguish those buildings which had collapsed or were partially collapsed. Thus the assignment of grade 5 or grade 4 could 36

37 be made with some confidence, though the assignment of grade 5 rather than grade 4 was sometimes a matter of judgement. For a number of buildings, damage at grade 3 or 2 could be clearly seen (local wall or roof failures, parapet failures), etc. Table 4 shows the comparison of the results from ImageCat-GEOCAN and Pictometry. The overall accuracy (assuming Pictometry is 100% accurate) is 79.5%. The table shows that about 80% of the grade 5 buildings in Pictometry were identified as grade 4 or grade 5 in GEOCAN, which is encouraging. Likewise, approximately 80% of the GEOCAN grade 5 was identified as grade 4 or grade 5 in Pictometry. However, when we look at grade 4 alone in Pictometry, 60% of the Pictometry grade 4 data were omitted in GEOCAN. It is clear that increasing the accuracy of the grade 4 would be an important step forward. Even with Pictometry data, the interpreters encountered difficulties identifying grade 3 buildings. A more detailed account of the analysis undertaken by CAR can be found in Spence and Saito (2010). The Pictometry damage assessment was subsequently validated using ground survey results collected by the Earthquake Engineering Field Investigation Team (EEFIT). The overall accuracy of the Pictometry assessment when compared against the ground survey results is 74%. Booth et al. (in review) describes the assessment in more detail. This need for a cascaded accuracy assessment approach from nadir, to off-nadir, to field assessment was necessary because of immediate access issues by ground teams and the need to develop rapid assessments of damage that could be improved over time. 37

38 Table 4: Comparison between the GEOCAN and Pictometry Damage Assessment Results Aerial interpretation by GEOCAN 0 D4 D5 Total Us er s accuracy Commission error Pictometry interpretation nvd+d2+ D % 10% D % 80% D % 36% Total Producers accuracy 88% 23% 68% Omission error 12% 77% 32% Kappa value: 0.48 The Pictometry data proved to be useful and the only source of validation data during the period where no ground damage data were available. Pictometry assessment results were subsequently compared against ground survey data by EEFIT (Booth et al., forthcoming) which demonstrated that even Pictometry has the tendency to underestimate the damage. A Bayesian updating method is being proposed that updates the remote sensing damage assessment results by incorporating ground data as they become available (Booth et al., forthcoming). The ground survey validation by WB-JRC- 38

39 UNOSAT has identified three areas that need to be improved in order for the accuracy of the damage assessment to increase: 1) create reliable baseline data (building points/footprints), 2) reduce interpreter errors, and 3) reconsider the damage grades to be used for remote sensing based damage assessment. WB-UNOSAT-JRC Damage Totals The results of the joint WB-UNOSAT-JRC damage analysis (Table 5) shows that a little over 90,000 buildings were either destroyed or experienced heavy damage in the earthquake (damage grades 3 through 5). This represents a little less than 1/3 of the building inventory in the affected areas. Most of the damage occurred in the Port-au- Prince area; however, significant numbers of buildings were also destroyed in Carrefour, Delmas, Leogane and Petion-Ville. Based on median floor area estimates for different occupancy uses, this damage translates roughly to over 26 million square meters in building area affected with about a third of this total associated with buildings that will have to be either replaced or significantly repaired. The total repair cost to buildings is estimated by the joint DaLA report (Kemper et al., 2010) to be over US$6 billion. Table 5: Number of Damaged Houses Grouped in EMS-98 Damage Grades per Commune and Dominant Land-Use Class EMS-98 Damage Grades COMMUNE Carrefour Cite Soleil

40 EMS-98 Damage Grades COMMUNE Delmas Grand-Goave Gressier Jacmel Leogane Petion-Ville Petit-Goave Port-Au-Prince Tabarre Total The WB-UNOSAT-JRC team delivered its findings to the DaLA team well in time for the New York Donor Conference held on 31 March Because of the tight timeframe associated with this delivery, the damage results above were considered preliminary. The Ministry of Public Works of Haiti, with support from and co-financed by the World Bank, the Global Facility for Disaster Reduction and Recovery (GFDRR) and the United States Agency for International Development (USAID) Office of Foreign Disaster Assistance (OFDA), has conducted a house-to-house assessment of building damage levels in order to ensure the safety of each building and to estimate the rebuilding or 40

41 repair costs of all affected structures 10. This level of investigation is considered the most reliable and most accurate indicator of damage and thus represents the benchmark to which the remote-sensing assessment must be compared. A final comparison is not possible at this stage; however, early indications are that the remote-sensing based results are very reasonable with respect to counts of heavily-damaged or destroyed buildings. Damage Maps and Other Products A number of products were produced by the joint effort. These included Atlas Sheets that showed the location of damaged buildings, damage maps that described the state of damage of individual buildings, and vector-based maps that showed damage to critical roadways or other linear infrastructure (see sample data in Plate 4). These products are described below. Atlas Sheets and Other Products - In addition to the comprehensive joint damage assessment that was fed into the PDNA process, a series of 11 atlases were produced covering the most affected areas. Individual atlases consisted of multiple map sheets, each of them showing the latest damage assessment and baseline data, such as road names from Open Street Map (OSM), available at the time. Damage representation in the atlas series was consistent with that for the PDNA, using EMS 98. Individual map sheets were organized using a grid for each city with corresponding map sheet numbers. The atlas could be printed as a whole, or only for the relevant pages of specific areas of 10 Note that the WB/GFDRR financed portion of the structural assessment is being implemented by UNOPS and the USAID OFDA portion is being implemented by the Pan American Development Foundation (PADF). 41

42 interest. In order to be of use during field assessments and for easier validation, a scale of 1:2,500 for A3 size hardcopy prints was selected. If printed in A4 size, the map scale was 1:3,500. Despite the ortho-rectification, the horizontal displacement error was approximately ± 5 m. Due to the large file size of the atlases, ranging from 25 to 600MB, digital download from Haiti was challenging. Hardcopies were distributed in Port-au- Prince to relevant organizations, as well as through CNIGS. Atlases for the following locations were produced: Port-au-Prince, Leogane, Carrefour, Jacmel, Delmas, Petion Ville, Grand Goave, Petit Goave, Cite Soleil, Tabarre and Gressier. Swisstopo and the Remote Sensing Laboratories (RSL) at the University of Zurich provided extensive pre-event building data in direct technical support of the joint damage analysis that was incorporated into the atlases and the PDNA assessment. Initial reference data were received from collaborative centers (JRC with the support of the Stability Instrument; Sertit/Space Charter) and other supporting entities (ITHACA 11, DLR 12, OSM). Multiple damage maps were created by various providers during the Haiti crisis. In fact, there were so many maps created that it became a problem for many of the users in the field. Although not all were building damage maps, there were days when more than 120 maps were produced. In the first weeks following the earthquake, over 2,000 maps were produced. This caused serious information overload and UNOSAT is currently working with the Global Disaster Alert and Coordination System (GDACS), 11 Information Technology for Humanitarian Assistance, Cooperation and Action 12 German Aerospace Center 42

43 coordinated by the Office for the Coordination of Humanitarian Affairs (OCHA) and with input from JRC and numerous country stakeholders, to better organize the information system so that early responders will be better able to select relevant maps. This is being done through cataloging maps on the Virtual On Site Operations Coordination Centre (VirtualOSOCC). Although the direct benefit of the building damage maps was limited for the search and rescue teams on the ground, these maps were used for early recovery efforts by UNDP 13, as well as feeding into the PDNA assessment. One issue that caused some confusion and initial inconsistencies was the simple question of what constitutes a building. Cases where a building consists of more than one building (e.g., building wing), represented as one point or multiple points for complex and/or large structures, affected the damage map. To overcome this issue, an interpolated building damage intensity surface covering downtown Port-au-Prince was produced, as assessed from both satellite and aerial survey imagery. This weighted the intensity to allow for a more balanced visual representation of downtown areas with damaged large governmental structures and densely populated shanty areas. For the joint methodology, the GEOCAN building footprint polygons were converted to point data to be consistent with the work of EC JRC and UNOSAT, thus facilitating a homogenous dataset. Vector data sharing - The situation was complex also due to the number of aid or relief actors on the ground. In order to move towards common data or formats on which to base decisions, remote sensing derived vector data were shared with actors on the 13 United Nations Development Program 43

44 ground and at headquarters. This included initial damage assessments of, for example, governmental buildings, but also more dynamic parameters, such as status of road obstacles and location/gatherings of internally displaced people (IDPs). Vectors were shared as klm files, ESRI shapefiles, and ESRI GeoDatabase to meet the various needs by non-remote-sensing expert mapping entities. CNIGS played an important role in sharing data locally and ensuring inclusion in the Government assessments and subsequent reconstruction efforts. Additionally, UNOSAT consolidated the different building damage datasets produced by the three organizations (WB-UNOSAT-JRC) into one ESRI geo-database. A dedicated web page for downloading these data was created by UNITAR/UNOSAT. These products are available in different GIS formats (.shp and.gdb) and can be downloaded at 44

45 Plate 4: Sample Product from Joint WB-UNOSAT-JRC Collaboration 45